Fault detection and accommodation by means of neural networks. Application to the boiler unit Krzysztof Patan ∗,1 J´ozefKorbicz ∗ ∗ Institute of Control and Computation Engineering University of Zielona G´ ora ul. Podg´orna 50, 65-246 Zielona G´ ora, Poland e-mail: {k.patan,j.korbicz}@issi.uz.zgora.pl Abstract: The work presented in this paper deals with a fault tolerant control system designed for a boiler unit. The main core of the proposed system is the so-called on-line fault approximator built using locally recurrent neural networks. The on-line stable training of the fault approximator is developed for monitoring of the controlled system. The obtained fault estimator is then used for the fault detection as well as for the fault accommodation. Computer experiments illustrate the performance of the proposed system for a boiler unit. Copyright c  2009 IFAC Keywords: Neural networks, Fault detection, Fault accommodation, Fault tolerant control, Boiler unit. 1. INTRODUCTION The increasing requirements for high levels of system per- formance and reliability in the presence of unexpected changes of system functions cause that Fault Tolerant Con- trol (FTC) systems have received the increasing attention in the last years Blanke et al. [2006]. Sensor or actuator faults, product changes, the material consumption may affect the controller performance Korbicz et al. [2004], Blanke et al. [2006]. The main objective of a FTC system is to maintain the current performance of the system as close as possible to the desirable one, and preserve sta- bility conditions in the presence of faults. To date, many different FTC schemes were investigated Patton [1997], Blanke et al. [2006], Zhang [2007]. The existing FTC methods can be divided into two groups: passive and active approaches Zhang [2007]. Passive aproaches are designed to work with a presumed failure modes and its perfor- mance tends to be conservative, especially in the case of unanticipated faults. In contrast, active methods reacts to the occurrence of system faults on-line and attempt to maintain the overal system stability and performance even in the case of unanticipated faults. This paper presents an active approach for designing an automated fault de- tection and accommodation system. Firstly, the detection of faults is performed on-line, and then accommodation is carried out to self-correct a particular failure through the reconfiguration of the control system. Both actions are carried out using the so-called on-line fault approximator Polycarpou and Vemuri [1995]. In this work we propose to construct the on-line fault approximator using locally recurrent neural networks Patan [2008]. This is motivated 1 This work was supported in part by the Ministry of Science and Higher Education in Poland under the grants N N514 1219 33 and R01 012 02 (DIASTER). by the fact that this class of neural networks can model wide class of dynamic processes simultaneously possessing relatively simple neuron interconnection structure, which makes it possible to derive stable training algorithms rel- atively easily Patan [2007]. The performance of the pro- posed fault detection and accommodation architecture is tested on a boiler unit. All experiments are carried out in Matlab/Simulink environment. 2. FAULT DETECTION AND ACCOMMODATION Consider a nonlinear dynamic system governed by the following equation: x(k + 1) = g ( x(k), u(k) ) + f ( x(k), u(k) ) , (1) where g is a process working at the normal operating conditions, x(k) is the state vector, u(k) is the control input vector and f represents a fault affecting the process. The unknown function f is a deviation in the system caused by a fault. In the case considered, f is a function of both the state x and control input u, which makes it possible to model wide class of possible faults, not only the additive ones Polycarpou and Vemuri [1995]. To model deviations in system dynamics due to a fault, let introduce the nonlinear estimator in the form: ˆ x(k + 1) = ˆ g ( x(k), u(k) ) + ˆ f ( x(k), u(k); ˆ θ(k) ) +G(x(k) − ˆ x(k)), (2) where ˆ g is a model of the process, ˆ f is the approxima- tion of the fault, ˆ θ is a vector of adaptable parameters and G is a constant stable matrix. The initial value of adaptable parameters should be selected such that ˆ f ( x(k), u(k); ˆ θ(k) ) = 0 when the system works at the Proceedings of the 7th IFAC Symposium on Fault Detection, Supervision and Safety of Technical Processes Barcelona, Spain, June 30 - July 3, 2009 978-3-902661-46-3/09/$20.00 © 2009 IFAC 119 10.3182/20090630-4-ES-2003.0188